Re: 2.4.10aa1 - 0-order allocation failed.

[Andrea Arcangeli]

On Wed, Sep 26, 2001 at 12:02:05PM -0300, Marcelo Tosatti wrote:
> 
> 
> On Wed, 26 Sep 2001, Andrea Arcangeli wrote:
> 
> > On Wed, Sep 26, 2001 at 06:07:48PM +0400, Oleg A. Yurlov wrote:
> > > 
> > >         Hi, Andrea,
> > > 
> > >         We have next problem on our servers:
> > > 
> > > Sep 26 11:22:39 sol kernel: __alloc_pages: 0-order allocation failed (gfp=0x20/0)
> > > Sep 26 11:22:39 sol kernel: f048dd94 e02ab000 00000000 00000020 00000000 00000020 00000020 e298f820 
> > > Sep 26 11:22:39 sol kernel:        e298f844 00000001 e030a56c e030a6c4 00000020 00000000 e01382be 00000000 
> > > Sep 26 11:22:39 sol kernel:        e013874a e013488c 00000000 e298f820 00000202 e298f898 00000202 00000246 
> > > Sep 26 11:22:39 sol kernel: Call Trace: [put_dirty_page+122/132] [flush_old_exec+234/572] [sys_ustat+212/268] [kill_super+232/352] [unix_gc+394/748] 
> > > Sep 26 11:22:39 sol kernel:    [Unused_offset+27374/99203] [Unused_offset+12842/99203] [call_spurious_interrupt+14521/27705] [Unused_offset+43342/99203] [call_spurious_interrupt+14615/27705] [call_spurious_interrupt+16483/27705] 
> > > Sep 26 11:22:39 sol kernel:    [Unused_offset+90704/99203] [ipgre_rcv+233/636] [ipgre_rcv+503/636] [fcntl_getlk+327/624] [do_invalid_TSS+43/96] 
> > > Sep 26 11:22:39 sol kernel: __alloc_pages: 0-order allocation failed (gfp=0x20/0)
> > > Sep 26 11:22:39 sol kernel: f048ddd4 e02ab000 00000000 00000020 00000000 00000020 00000020 e298f820 
> > > Sep 26 11:22:39 sol kernel:        e298f844 00000001 e030a56c e030a6c4 00000020 00000000 e01382be 00000000 
> > > Sep 26 11:22:39 sol kernel:        e013874a e013488c 00000000 e298f820 00000202 e298f898 00000202 00000246 
> > > Sep 26 11:22:39 sol kernel: Call Trace: [put_dirty_page+122/132] [flush_old_exec+234/572] [sys_ustat+212/268] [kill_super+232/352] [unix_gc+394/748] 
> > > Sep 26 11:22:39 sol kernel:    [Unused_offset+27374/99203] [call_spurious_interrupt+13905/27705] [call_spurious_interrupt+17048/27705] [Unused_offset+90704/99203] [ipgre_rcv+233/636] [ipgre_rcv+503/636] 
> > > Sep 26 11:22:39 sol kernel:    [fcntl_getlk+327/624] [do_invalid_TSS+43/96] 
> > 
> > the system.map is wrong but this should be harmless, just a notice (if
> > you do the reverse lookup to find the address and you resolve the right
> > symbols we could make sure of that).
> > 
> > For driver writers (since it could be on topic with those GFP_ATOMIC
> > faliures): as I suggested to the SG folks make sure to never use
> > GFP_ATOMIC in normal kernel context, if you want lowlatency use GFP_NOIO
> > instead. GFP_NOIO can schedule (so you must release all the spinlocks
> > first) but it will never block on I/O so it will provide a small latency
> > too _but_ it will be able to shrink the clean cache so it is very unlikely
> > it will fail unless you have lots of dirty or mapped cache in ram.
> > 
> > >         Also, we see next in process status:
> > > 
> > > USER       PID %CPU %MEM   VSZ  RSS TTY      STAT START   TIME COMMAND
> > > vz         927  0.0 625.1 43900 4267034752 ? S    08:10   0:00 hits
> > > vz        1030  0.0 625.1 43900 4267034752 ? S    08:11   0:00 hits
> > > vz        4561  1.3 625.1 45948 4267034724 ? S    10:48   0:00 hits
> > > root      4564  0.0  0.0  1460  548 pts/2    S    10:48   0:00 grep hits
> > > vz        4566  0.0 625.1 45948 4267034724 ? S    10:48   0:00 hits
> > 
> > Ben sent the fix for this one [Linus, you can find it on l-k if you
> > weren't cc'ed] (was a missing check in the tlb shootdown smp fixes) but
> > it's only a beauty issue, so really don't worry about it :)
> > 
> > >         After these errors we see some uninterruptable processes (with flag D in
> > > process  status),  gdb  say  that function "fdatasync" called and no returned...
> > > Soft reboot not work.
> > > 
> > >         Server   has   2  CPUs (Pentium III Katmai), 2Gb RAM, 2Gb swap, Hardware
> > > RAID (Mylex DAC960PTL1 PCI RAID Controller).
> > > 
> > >         Any ideas ?
> > 
> > Yes you have highmem.
> > 
> > Last night I spent one hour on the traces from Bob (btw, many thanks for
> > the helpful report Bob!) and the first suspect is the recent
> > GFP_NOHIGHIO logic.
> > 
> > Despite Bob's traces not obviously showing this, I think I can see a
> > potential problem with writepage with regard to the GFP_NOHIGHIO logic
> > (I just checked 2.4.9ac15 has the same issue too, see the CAN_DO_FS
> > definition so this shouldn't been introduced recently).
> > 
> > This should fix it, and please also apply vm-tweaks-2 posted to l-k a
> > few minutes ago.
> > 
> > --- 2.4.10aa1/mm/vmscan.c	Sun Sep 23 22:16:22 2001
> > +++ vm/mm/vmscan.c	Wed Sep 26 16:34:30 2001
> > @@ -392,7 +384,7 @@
> >  			int (*writepage)(struct page *);
> >  
> >  			writepage = page->mapping->a_ops->writepage;
> > -			if ((gfp_mask & __GFP_FS) && writepage) {
> > +			if ((gfp_mask & __GFP_FS) && ((gfp_mask & __GFP_HIGHIO) || !PageHighMem(page)) && writepage) {
> >  				ClearPageDirty(page);
> >  				page_cache_get(page);
> >  				spin_unlock(&pagemap_lru_lock);
> > 
> > 
> > And if the above patch still doesn't help can you just apply this below
> > patch to disable the NOHIGHIO logic all together, just to make sure
> > we're looking in the right place?
> > 
> > --- 2.4.10aa1/mm/highmem.c.~1~	Sun Sep 23 21:11:43 2001
> > +++ 2.4.10aa1/mm/highmem.c	Wed Sep 26 16:38:34 2001
> > @@ -328,7 +328,7 @@
> >  	struct page *page;
> >  
> >  repeat_alloc:
> > -	page = alloc_page(GFP_NOHIGHIO);
> > +	page = alloc_page(GFP_NOIO);
> >  	if (page)
> >  		return page;
> >  	/*
> > @@ -366,7 +366,7 @@
> >  	struct buffer_head *bh;
> >  
> >  repeat_alloc:
> > -	bh = kmem_cache_alloc(bh_cachep, SLAB_NOHIGHIO);
> > +	bh = kmem_cache_alloc(bh_cachep, SLAB_NOIO);
> >  	if (bh)
> >  		return bh;
> >  	/*
> > 
> > Of course also make sure that a SYSRQ+e or SYSRQ+i doesn't relieve the
> > machine and allows to kill the D tasks :).
> 
> Andrea, 
> 
> I don't understand why you "removed" the SLAB_NOHIGHIO flag from the
> bounce buffering allocation and used SLAB_NOIO instead.

In case it wasn't clear enough in my last email I didn't removed
anything in my tree, the above is a test patch just in case the
writepage fix isn't enough to cure the deadlock. If the deadlock goes
away with the above patch we know the bug is in the nohighio logic.

> 
> The SLAB_NOHIGHIO flag allows the bounce buffering code to block on lowmem
> writes (which is allowed), thus avoiding allocation failures.
> 

Andrea